Engineering Failure Analysis 59 (2016) 211222

Contents lists available at ScienceDirect

Engineering Failure Analysis

journal homepage: www.elsevier.com/locate/efa

Stochastic analysis on exural behavior of reinforced concrete

beams based on piecewise response surface scheme
Tengfei Xu a,b,c, Tianyu Xiang a, Renda Zhao a, Guotao Yang b, Cheng Yang a,c,
a
Department of Bridge Engineering, Southwest Jiaotong Unviersity, Chengdu 610031, PR China
b
Centre for Infrastructure Engineering and Safety, School of Civil and Environmental Engineering, The University of New South Wales, UNSW, Sydney, NSW 2052, Australia
c
Key Laboratory of High-speed Railway Engineering, Ministry of Education, Southwest Jiaotong University, Chengdu 610031, PR China

a r t i c l e i n f o a b s t r a c t

Article history: To save the computational efforts of Monte Carlo simulation together with nonlinear nite
Received 23 March 2015 element method, the analysis framework combining the response surface method and Monte
Received in revised form 22 September 2015 Carlo simulation is usually adopted to investigate the stochastic nonlinear behavior of
Accepted 5 October 2015
structures. It is found that the traditional response surface method could not describe stochas-
Available online 14 October 2015
tic behavior of cracked concrete beams. In order to overcome the discontinuity caused by
cracking nonlinearity, a scheme by introducing a piecewise response surface is proposed in
Keywords: this paper. This scheme is evaluated by the stochastic analysis of several concrete beams. The
Reinforced concrete
comparison between the proposed method and some traditional methods, including Monte
Deection
Carlo simulation and Monte Carlo simulation with traditional response surface, shows that
Random analysis
Response surface method the proposed method could well depict the stochastic behavior of concrete beams. Finally,
Creep and shrinkage the probability density evolution of concrete beams from short-term deection to a long-
term deection is analyzed.
2015 Elsevier Ltd. All rights reserved.

1. Introduction

To ensure the serviceability of reinforced concrete structures, deection control is very important in engineering practice, as
excessive deformation is one of the most common causes of damage and could result in large annual cost to the construction in-
dustry [1]. Actually, both overestimating and underestimating deformation of the engineering structures could cause safety prob-
lem in engineering. For example, in order to maintain a satised elevation of a bridge, pre-camber should be set based on the
deection prediction under long-term dead loads and live loads. Either overestimating or underestimating the pre-camber of a
bridge will lead to track irregularity in high-speed railway, which may result in discomfort to passengers and insecurity to trains.
Therefore, predicting the deformation of concrete structure as exactly as possible is essential to the industry community.
Due to the natural uncertainties of mechanical characters of concrete [2], such as the modulus of elasticity, the compressive
and tensile strength, and the creep and shrinkage parameters, a deterministic analysis approach for calculating the deections
of a concrete structure is likely difcult to predict the experimental-measured deections. In 1979, Ramsay et al. [3] studied
the statistical variation of the immediate deection of reinforced concrete (RC) beams, and the obtained coefcient of variation
was between 25% and 30%. Creep and shrinkage is another uncertainty material property of concrete structures [4,5]. In order
to take into account the randomness, several numerical methods has been adopted by the researcher to investigate the stochastic
response of concrete structures, such as Monte Carlo Simulation (MCS) method [2,3,6,7] and Latin Hypercube Simulation (LHS)
method [5,8].

Corresponding author at: Department of Bridge Engineering, Southwest Jiaotong Unviersity, Chengdu 610031, PR China.
E-mail address: yangcheng@swjtu.edu.cn (C. Yang).

http://dx.doi.org/10.1016/j.engfailanal.2015.10.004
1350-6307/ 2015 Elsevier Ltd. All rights reserved.
212 T. Xu et al. / Engineering Failure Analysis 59 (2016) 211222

Fig. 1. Stressstrain relationship for concrete.

Monte Carlo Simulation method, which adopts a random sampling technique based on a computer-based analytical or numer-
ical model, could provide the most reliable result in simulating the stochastic behavior of a structural system [9]. However, the
MCS method takes huge computational cost, especially when the nonlinear nite element analysis is required in every sampling.
Instead, the response surface method (RSM), which approximates the implicit response function by a simple and explicit polyno-
mial, provides an effective and versatile way to applying numerical methods [9,10] The computing cost could be saved signicant-
ly in MCS if the structure response could be depicted by RSM.
It is worth noting that the major cause of the variability of the deection is the variability of the beam stiffness, which is af-
fected by cracking of the concrete and the variability in the concrete strength, particularly the tensile strength [3]. Due to the large
discrepancy between cracked and uncracked stiffness [2], the real deection response surface of a RC beam with respect to the
tensile strength of the concrete may be unsmooth and convex, which could lead to the failure in estimating the random deec-
tions if regular response surface method is adopted. To this end, a piecewise response surface method is proposed in this paper to
solve the aforementioned problem, which is not reported yet in literature. By using the proposed method, the evolution of imme-
diate deection to long-term deection of the test beams is also investigated.

2. Stochastic analysis model

2.1. Finite element formulations

The nonlinear nite element analysis should be carried out to reect the actual loading-deection behavior of concrete structures. A
general nite element analysis program with degenerated beam element is adopted for immediate and long-term analysis [11]. The the-
ory behind the degeneration technique has been widely used in the formulation of shell and plate element [12,13]. Based on the theory of
Timoshenko beam, the three-dimensional displacement eld of degenerated beam is described by the nodal variables. Meanwhile, the
straindisplacement equations of three-dimensional structure are retained. To integrate the stiffness matrix and internal force, the piece-
wise integration method is usually used in the degenerated beam element [12,13]. In this method, the element is subdivided into several
pieces, and one integration point is attributed for every piece. The stress of integration point can be computed following one-dimensional
constitutive model. With this approach, the material nonlinearity of steel and concrete can be simulated conveniently. In the present
study, the one-dimensional perfect elasto-plastic constitutive model is adopted for steel. Meanwhile, a nonlinear strainstress relation-
ship for concrete shown in Fig. 1 is introduced in the present study.

Table 1
Statistical properties of random variables.

Variables Distribution type Mean Coefcient of variation References

1 Normal 1.0 0.26 [5]

2 Normal 1.0 0.25 [5]
3 Normal 1.0 0.15 [21]
4 Normal 1.0 0.20 [4]
5 Normal 1.0 0.20 [22]
6 Normal 0.3 0.15 [20]
 T. Xu et al. / Engineering Failure Analysis 59 (2016) 211222 213

Fig. 2. Experimental points of RSM for bivariate function. (a) Experimental points of traditional RSM. (b) Experimental points of piecewise RSM.

2.2. Creep and shrinkage model

Several creep and shrinkage models, such as ACI 209 model [14], CEB-FIP MC90 model [15], B3 model [16], GL2000 model
[17,18], and so on, are proposed in those previous researches. In the present study, GL2000 model is chosen to simulate the
creep and shrinkage of concrete. The compliance function J(t, t0), which represents the stain at age t caused by a unit uniaxial
stress acted at age t0, is dened in GL2000 model as.

1 1
J t; t 0 t; t 0 1
Ec t 0 Ec

where Ec(t0) and Ec are elastic modulus of concrete at age t0 and 28 days, respectively. (t,t0) is creep coefcient, which is given as.
( " # ss s)
tt 0 0:3 7 tt 0   tt 0
2
t; t 0 t c 2 2:5 11:086h 2
tt 0 0:3 14 t 0 tt 0 7 tt 0 0:15V=S2

where V/S represents volumesurface ratio (mm), V is the volume of concrete, and S is the area of concrete surface exposing to
the air; h stands for the relative humidity (R.H.). The rst and second terms in brace of right hand of Eq. (2) represent basic creep,
and the third term is corresponding to drying creep. (tc)is expressed as.
8
>
> 1 t 0 t c
<vu

s
t c tu t 0 t c 3
>
> 1   t 0 Nt c
: t 0 t c 0:15 VS 2

Fig. 3. Traditional response surface for piecewise function with a single parameter.
214 T. Xu et al. / Engineering Failure Analysis 59 (2016) 211222

Fig. 4. Piecewise response surface for piecewise function with a single parameter.

in which tc is age of drying beginning. When only basic creep is concerned, (tc) =1.0 is set to be 1.0.
Shrinkage can be estimated with the following equations in GL2000 model.

sh shu ht 4

where.
s
30
shu 0:0009K 5
f cm28

4
h 11:18h 6

s
tt c
t   7
tt c 0:15 VS 2

in which K is a coefcient dependent on cement type [18].

2.3. Sources of uncertainty and variability

Various parameters may contribute to statistical scatter of creep and shrinkage effect, and model uncertainty of creep and
shrinkage prediction is considered to be one of the most important resources [5]. Generally, the stochastic model for time-
variant strain can be written as.
Z t
t; t 0 1 J t; t 0  t 0 1 J t; d 2 sh t 8
t0

in which 1 and 2 are the uncertain parameters related to creep and shrinkage model; stands for the stress of concrete.

Fig. 5. Cross-sections of beams.

 T. Xu et al. / Engineering Failure Analysis 59 (2016) 211222 215

Table 2
Section parameters of reinforced concrete beams.

No. As(mm2) d(mm) A's(mm2)

B1 570 213 63
B2 570 213 63
B3 787 205 63

The uncertainties of model parameters should also be concerned in the stochastic analysis. For GL2000 model, the model pa-
rameters that possess random nature include mean compressive strength of concrete fcm and mean ambient humidity h. The ran-
domness of mean compressive strength of concrete and ambient of humidity is expressed as 3fcm and 4h.
Furthermore, the randomness of elastic modulus of concrete must be considered. As shown in Eq. (1), for instantaneous elastic
strain computation, the elastic modulus at age t0 is used. Meanwhile, the elastic modulus at age 28 days is adopted in the second
term of left hand of Eq. (1). Generally, it is accepted that strong correlation may exist between elastic modulus of concrete and its
compressive strength. Keitel and Osburg [19] involve correlation between elastic modulus of concrete and its compressive
strength in uncertain analysis of creep and shrinkage effect with a correlation coefcient between those two parameters. An
alternative approach will be adopted in the present research. Through multiplying the deterministic development function of
elastic modulus with a random variable, the stochastic correlation between time-variant elastic modulus of concrete and its
mean compressive strength at 28 days for GL2000 model is dened respectively as.
0 s1
t 0:75 A
Ec t 5 @3500 4300 f cm 9
a bt 0:75

where the unit of fcm is MPa, s, a and b are constants related to cement type [18].
Although the tensile strength of concrete does not affect the strength of reinforced concrete structures, it should not be
neglected in deection calculations due to its inuence on the overall stiffness of beam. In this model, tensile strength of concrete
ft can be expressed via fcm as [20].

2=3
f t 6 f cm 10

Finally, the stochastic model for long-term strain can be nally written as.
 
1 t; t 0 ; 3 f cm ; 4 h
t; t 0 t 0  1
5 Ec 3 f cm ; t 0 5 Ec 3 f cm
Z t  
1 t; ; 3 f cm ; 4 h 11
1 d
t0 5 Ec 3 f cm ; 5 Ec 3 f cm
2 sh t; 3 f cm ; 4 h

The statistical parameters of the random variables are summarized in Table 1. Especially, when the immediate deection is
concerned, it can be found that only 3, 5, and 6 should be involved in the stochastic analysis.

3. Stochastic analysis method

3.1. Piecewise response surface method

Though approximating the implicit structural response with an explicit form, response surface approach provides a versatile
way to improve computation efciency in random analysis [23]. For engineering application, a simple form of response with a
second-order polynomial is usually used [24].

X
n X
n
2
g x 0 i x i ni xi 12
i1 i1

Table 3
Loading parameters of test beams.

No. l(mm) fc(MPa) Md(N .m) t(days) Loading ratio ()

B1 6706 45.7 16098 785 10.42%

B2 6706 45.1 16098 817 11.40%
B3 7925 29.9 22514 813 104.91%
216 T. Xu et al. / Engineering Failure Analysis 59 (2016) 211222

Fig. 6. Probability density curves of immediate deection calculated by MC, RSM + MC, and Piecewise-RSM + MC (B1 and B2).

in which n is the number of random variables, and 0...2n are the 2n + 1 unknown coefcients to be determined. A common
approach is to simulate g(x) at 2n + 1 experimental combinations of i and i + i. The unknown coefcients can be obtained
by

X g 13a
2 2 2 2
1 1 i n 1 i n
6 2 2 2
6 1 1 1 i n 1 1 i n
6
6 1 i n 1 1
2

2
i
2
n
6 1 1
6
6
X6 61 1 i i n
2
1 i i
2

2
n 13b
6 2 2 2
61 1 i i n 1 i i n
6
6
6 2 2 2
41 1 i n n 1 i n n
2 2 2
1 1 i n n 1 i n n

T
0 ; ; 2n 13c

T
g fg 0 ; ; g 2n g 13d

where i is the mean value of xi, i is the standard deviation of xi, is the arbitrary coefcient, and gi is the simulated result in
each experiment. Thus, the values of = 2 and n = 6 are used in the paper. A simple case of experimental points for bivariate
function by traditional RSM is plotting in Fig. 2(a).

Fig. 7. Probability density curves of immediate deection calculated by MC, RSM + MC, and Piecewise-RSM + MC (Beam B3).
 T. Xu et al. / Engineering Failure Analysis 59 (2016) 211222 217

Fig. 8. Immediate deection of B2 calculated by direct nonlinear FEM simulation, RSM simulation, and Piecewise-RSM simulation.

Generally, the response surface method can reduce the time-consuming computational effort while maintaining the proper ac-
curacy. But it is difcult to ensure the accuracy of traditional smooth response surface, when the real response curve is unsmooth
and convex, e.g. the piecewise function with a single parameter shown in Fig. 3. Similarly, in this study, the concrete cracking
strength, which is the most pronounced property, affecting serviceability calculations would lead to unsmooth and convex re-
sponse deection of reinforced concrete beam [25].
To solve this problem, a piecewise response surface method is suggested. The process is as follows:
Selecting the key parameter, which induces the unsmooth response. (Without loss of generality, x1 is labeled as the key
parameter for example).
Dividing the domain of key parameter x1 into some equal sub-intervals.
Uniting with other parameters (xi,i = 2 n), and modifying the matrix in Eq. (13b) as Eq. (14).
Constructing response surfaces in each sub-interval of key parameter as the same as traditional method.
Combining the response surfaces in each sub-interval, a complete response surface can be gained.
2  2
j j 2 2
1 m i n m i n
6  2
6
61 s
j
i n s
j

2
i
2
n
6  2
6
61 e
j
i n e
j

2
i
2
n
6
6
6  2
6 j j 2 2
X61 m i i n m i i n 14
6  2
6
61 m
j
i i n m
j
i i
2

2
n
6
6
6  2
61 j
i n n
j

2
i n n
2
6 m m
4  2
j j 2 2
1 m i n n m i n n

in which m j, sj and e j are the mid-point, the start point and the end point of the jthsub-interval of key parameter.

Fig. 9. Immediate deection of B3 calculated by direct nonlinear FEM simulation, RSM simulation, and Piecewise-RSM simulation.
218 T. Xu et al. / Engineering Failure Analysis 59 (2016) 211222

Fig. 10. Probability density curves of long-term deection calculated by MC, RSM + MC, and Piecewise-RSM + MC.

A sampling case for the bivariate function (xi and xj) is shown in Fig. 2(b). xi is selected as the key parameter in this case. The
domain of xi is divided into a total of n parts. In each part, response surface is constructed as.
k k k k 2 k 2
g x 0 1 xi 2 x j 3 xi 4 x j k 1 ::n 15

To calculate the coefcients k0 . .k4 in Eq. (15), ve sampling points should be selected as.
Point 1: xi = mk, xj =2.
 T. Xu et al. / Engineering Failure Analysis 59 (2016) 211222 219

Fig. 11. Long-term deection of B3 calculated by direct nonlinear FEM simulation, RSM simulation, and Piecewise-RSM simulation.

Point 2: xi = ek, xj = 2.
Point 3: xi = sk, xj = 2.
Point 4: xi = mk, xj =2 + 2.
Point 5: xi = mk, xj =2 2.
By this way, the piecewise response surface is close to the real response curve of the single parameter case mentioned
above(Fig. 4). Then, Monte Carlo (MC) sampling will be carried out to calculate the random responses.
Since only 2n +1 times of nonlinear nite element analysis are involved in traditional forming of response surface, the compu-
tational cost will be saved dramatically. For the piecewise response surface form, the times of nite element analysis are
2mn + 1, in which m is dened as the number of sub-intervals of key parameter.

3.2. Sensitivity analysis

Sensitivity analysis in general is the study of how the output uncertainty of a model is apportioned to the model input param-
eter uncertainties. The sensitivity coefcient is.


gx 
xi
xi 
x
v
0
i i
12 16
u 
uX N 
u
t @g x x  A
j
j1
x j
x j j

The scaling by the standard deviation makes these measures dimensionless [26]. More importantly, it makes the parameter
variations equivalent in a statistical sense [26].

Fig. 12. Probability density evolution for Beam No. 2 along the loading time.
220 T. Xu et al. / Engineering Failure Analysis 59 (2016) 211222

Especially, to consider the different probability in each sub-interval about the key parameter of x1, the sensitive coefcient is
modied to.

gx 
xk
x1   
x1 1 h h1
v
0 12 P x1 ; x1 17
u 
uX N 
u
t @g x x  A
j
j1
x j
x j j

in which, P(xh1, xh+1

1 ) is the probability of x1 in the hth sub-interval between xh1 and xh+1
1 .

4. Numerical simulation

4.1. Reinforced concrete beams specimens

To verify the effectiveness of piecewise response surface method, a series of simply supported reinforced concrete beams are
selected as the numerical examples, which were collected in the long term sustained loading test data base [27].
The cross-sections of beams are shown in Fig. 5. The parameters of section details of beams are listed in Table 2. In Table 2, As
and A's are the areas of tensile and compressive reinforcement bar, respectively; and d is the effective depths of the cross-section.
Uniformly distributed loading was applied at 28 days and sustained for 785 to 813 days. The experiments were carried out in
open air, and a mean value 78% for R.H. was suggested [27].
In Table 3, l is the length of span of beam; fc is the average compressive strength of concrete at 28 days; Md is maximum bend-
ing moment due to dead load constituted by the beam's own weight and by the uniform sustained loading applied at the same
age. Especially, the loading ratio is dened to evaluate the loading level for test beams as the form as.

Ma M cr
18
Mcr

in which Ma is the applied moment, Mcr is the average cracking moment calculated with the average tensile strength of
concrete.

4.2. Probability density of short-term deection

The probability density curves of immediate deection for beams can be obtained by MC method, MC with traditional re-
sponse surface method and MC with piecewise response surface method. In each simulation by three methods, the times of
Monte Carlo sampling are 10,000.
In the random analysis, the B1 and B2 are treated as the same beams as the only difference between them is that the com-
pressive strength of B1 is higher 1.3%. The probability density curve of them is shown in Fig. 6. The experimental and calculated
deections of B1 and B2 are also plotted.
It can be found that the experimental deections are smaller than the predictions by ACI and CEB codes, and the deection of
B1 is much more nearly 50% than that of B2. It is an evidence of uncertainty of reinforced concrete behavior to demonstrate that
random analysis is essential to predict deection of reinforced concrete beam.
In addition, bi-modality of probability density curves simulated by direct MCS, which is considered as the exact result in ran-
dom analysis, can also be observed in Fig. 6. In Fig. 6, the result obtained by the present piecewise response surface method with
MCS agrees with the direct MCS result. In the bimodal distribution, the group of lower deections is associated with the samplings
which did not crack. Corresponding, the higher group is due to the cracking of samplings. In these cases, applied moments are
slightly larger than the average cracking moment (10.42% and 11.40%). Obviously, in deterministic simulation, these sections at
mid-spans are supposed to crack. Actually, it can be found Fig. 6 that the beams have a signicant probability to avoid cracking
due to the randomness of tensile strength of concrete.
In Fig. 6, it can be observed that MC method with traditional response surface failed to simulate the bi-modality of probability den-
sity. As we know, the success of RSM depends on the accuracy of tting results of response surface. To give a more clear insight about
this phenomenon, the tting results for the Beam B2 with traditional RSM and piecewise RSM in the dimension 6(random variable
tensile strength of concrete) are illustrated in Fig. 7. In this gure, 6 is in range of mean3 standard deviation. Meanwhile, the values
of 3 (random variable about compressive strength) and 5 (random variable about elastic modulus model) are considered as mean
values. The sensitive coefcient of immediate deection about 6 is also presented in the Fig. 7. It can be found that the deection
curve simulated by direct nonlinear nite element method is signicantly unsmooth. Obviously, the piecewise response surface sim-
ulation performs well, while a single smooth response surface cannot describe the real unsmooth curve of deection. The divergence
between results simulated by RSM and direct nonlinear nite element method, combing with high sensitive coefcient at the interval
of divergence, leads to the computational error of traditional RSM.
When the applied moment is signicantly larger than average cracking moment such as the B3, bimodal probability density is
not found in Fig. 8. For Beam B3, the curves calculated by MC with traditional response surface and with piecewise response
 T. Xu et al. / Engineering Failure Analysis 59 (2016) 211222 221

surface coincide well with those calculated by direct MC method. Fig. 9 shows that 6 is not sensitive to short-term deection in
this case, and the response curves of deection of beam B3 are nearly smooth by three methods.

4.3. Probability density of long-term deection

The probability density curves of long-term deection for test beams simulated by three methods are shown in Fig. 10.
Similarly with Fig. 8, results simulated by three methods for deection of the beam B2 under 6 ranged of mean3 standard
deviation are presented in Fig. 11. Meanwhile, the sensitive coefcient of 6 is also presented in Fig. 11.
In this case, the effect of creep and shrinkage might increase the risk of cracking, and decrease the uncertainty of cracking
behavior. For the reason, the sensitive coefcient of 6 in Fig. 11 is smaller than that in Fig. 11.
It is different from Fig. 6 that the bimodal distributions are not observed Fig. 10. Contrasting to Fig. 7, the divergence interval of
results simulated by RSM and direct nonlinear nite element method is with lower sensitive coefcient in Fig. 11, which would
explain the well-tting of results obtained by three methods in Fig. 10.
Owing to the reduction of computational costs by MC with Piecewise response surface method, probability density evolution
for beam No. 2 along the loading time from 0 to 55 days can be easily obtained in Fig. 12. It can be found that bimodal distribu-
tion transforms into unimodal distribution gradually, as the loading time goes on, because of the effect of creep and shrinkage.

5. Conclusions

Based on the nonlinear nite element model with degenerated beam element, a stochastic analysis method of reinforced
concrete beam is developed by combing MCS method with piecewise response surface method. The short-term and long-term
random deections taking into account cracking, creep and shrinkage can be calculated by present model. Compared with direct
MCS, this model could save the computational costs to a large extent. Besides, the piecewise response surface adopted in this
investigation could avoid the great discrepancy between a traditional response surface and real unsmooth response function,
and could bring about satised analysis results.
The proposed approach is used to evaluate variability of deections for a set of test beams. Bimodality of probability density
curves of immediate deections are observed, which is simulated successfully using direct MC method and MC method with
piecewise response surface. It is also found that with the development of creep and shrinkage, the risk of cracking increases.
The sensitivity analysis shows that the sensitive coefcient of tensile strength of the concrete to long-term deections is less
than that to the immediate deection, which gives rise to the evolution of the probability density curves from bimodal distribu-
tion to unimodal distribution gradually with the development of concrete age.

Acknowledgments

The authors are very grateful for the nancial support from the National Natural Science Foundation of China with Grant Nos.
51308468 and 51378432. Their thanks also go to the Fundamental Research Funds for the Central Universities with Grant No.
2682014CX072. The rst author and the fth author wish to thank the Key Laboratory of High-speed Railway Engineering, Ministry
of Education, Southwest Jiaotong University, People's Republic of China, for its support. Partial support was provided to the forth au-
thor by the Australian CSIRO Climate Adaptation Flagship through its Climate Adaptation Engineering for Extreme Events Cluster.

References

[1] A. Castel, R.I. Gilbert, G. Ranzi, Instantaneous stiffness of cracked reinforced concrete including steel-concrete interface damage and long-term effects, J. Struct.
Eng. 140 (6) (2014) (04014021).
[2] B.S. Choi, A. Scanlon, P.A. Johnson, Monte carlo simulation of immediate and time-dependent deflections of reinforced concrete beams and slabs, ACI Struct. J. 101
(5) (2004) 633641.
[3] R.J. Ramsay, S.A. Mirza, J.G. MacGregor, Monte carlo study of short time deflections of reinforced concrete beams, ACI Journal Proceedings, 76, ACI, 1979.
[4] H. Madsen, Z.P. Bazant, Uncertainty analysis of creep and shrinkage effects in concrete structures, ACI Journal Proceedings, 80, ACI, 1983.
[5] Z.P. Bazant, K.-L. Liu, Random creep and shrinkage in structures: ampling, J. Struct. Eng. 111 (5) (1985) 11131134.
[6] M.G. Stewart, D.V. Rosowsky, Time-dependent reliability of deteriorating reinforced concrete bridge decks, Struct. Saf. 20 (1) (1998) 91109.
[7] M.G. Stewart, Serviceability reliability analysis of reinforced concrete structures, J. Struct. Eng. 122 (7) (1996) 794803.
[8] A. Florian, An efficient sampling scheme: updated latin hypercube sampling, Probab. Eng. Mech. 7 (2) (1992) 123130.
[9] G. Yang, Research on fatigue of steel mid-pylon of three-pylon two-span suspension bridge under vehicle load(PhD Thesis ; Department of Bridge Engineering,
Tongji University; Shanghai, China) 2013.
[10] J. Cheng, C. Cai, R.C. Xiao, Probabilistic response analysis of cracked prestressed concrete beams, Adv. Struct. Eng. 10 (1) (2007) 110.
[11] T. Xiang, Y. Tong, R. Zhao, A general and versatile nonlinear analysis program for concrete bridge structure, Adv. Eng. Softw. 36 (10) (2005) 681690.
[12] S. Ahmad, B.M. Irons, O. Zienkiewicz, Analysis of thick and thin shell structures by curved finite elements, Int. J. Numer. Methods Eng. 2 (3) (1970) 419451.
[13] W. Kanok-nukulchai, A simple and efficient finite element for general shell analysis, Int. J. Numer. Methods Eng. 14 (2) (1979) 179200.
[14] Committee A. , Institute A.C. , for Standardization I.O., Building Code Requirements for Structural Concrete (aci 318-08) and Commentary, American Concrete
Institute, 2008.
[15] euro-international du bton C., CEB-FIP model code 1990: design code . 213-214, 1993 (Telford).
[16] Z.P. Baant, S. Baweja, Justification and refinements of model b3 for concrete creep and shrinkage 1. Statistics and sensitivity, Mater. Struct. 28 (7) (1995)
415430.
[17] N. Gardner, M. Lockman, Design provisions for drying shrinkage and creep of normal-strength concrete, ACI Mater. J. 98 (2001) 2.
[18] N. Gardner, Comparison of prediction provisions for drying shrinkage and creep of normal-strength concretes, Can. J. Civ. Eng. 31 (5) (2004) 767775.
222 T. Xu et al. / Engineering Failure Analysis 59 (2016) 211222

[19] H. Keitel, A. Dimmig-Osburg, Uncertainty and sensitivity analysis of creep models for uncorrelated and correlated input parameters, Eng. Struct. 32 (11) (2010)
37583767.
[20] D.V. Val, M.G. Stewart, R.E. Melchers, Effect of reinforcement corrosion on reliability of highway bridges, Eng. Struct. 20 (11) (1998) 10101019.
[21] A.S. Al-Harthy, B. Frangopol, Reliability-based design of prestressed concrete beams, J. Struct. Eng. 120 (11) (1994) 31563177.
[22] P. Laumet, Reliability-Based Deterioration Model for Deflection Limit State Girder Bridges, The University of Michigan, Michigan, 2006.
[23] J. Huh, A. Haldar, Stochastic finite-element-based seismic risk of nonlinear structures, J. Struct. Eng. 127 (3) (2001) 323329.
[24] T. Xu, T. Xiang, Y. Zhan, R. Zhao, Reliability analysis of circular concrete-filled steel tube with material and geometrical nonlinearity, J. Mod. Transp. 20 (3) (2012)
138147.
[25] J.J. Kim, M.M.R. Taha, T.J. Ross, Establishing concrete cracking strength interval using possibility theory with an application to predict the possible reinforced con-
crete deflection interval, Eng. Struct. 32 (11) (2010) 35923600.
[26] P.L. Liu, A.D. Kiureghian, Finite element reliability of geometrically nonlinear uncertain structures, J. Eng. Mech. 117 (8) (1991) 18061825.
[27] B. Espion, Long term sustained loading tests on reinforced concrete beams, Bulletin du Service Gnie Civil, 1988 (88-1).